1 /* $OpenBSD: uvm_pdaemon.c,v 1.64 2013/05/30 16:29:46 tedu Exp $ */ 2 /* $NetBSD: uvm_pdaemon.c,v 1.23 2000/08/20 10:24:14 bjh21 Exp $ */ 3 4 /* 5 * Copyright (c) 1997 Charles D. Cranor and Washington University. 6 * Copyright (c) 1991, 1993, The Regents of the University of California. 7 * 8 * All rights reserved. 9 * 10 * This code is derived from software contributed to Berkeley by 11 * The Mach Operating System project at Carnegie-Mellon University. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. All advertising materials mentioning features or use of this software 22 * must display the following acknowledgement: 23 * This product includes software developed by Charles D. Cranor, 24 * Washington University, the University of California, Berkeley and 25 * its contributors. 26 * 4. Neither the name of the University nor the names of its contributors 27 * may be used to endorse or promote products derived from this software 28 * without specific prior written permission. 29 * 30 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 31 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 32 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 33 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 34 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 35 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 36 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 37 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 38 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 39 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 40 * SUCH DAMAGE. 41 * 42 * @(#)vm_pageout.c 8.5 (Berkeley) 2/14/94 43 * from: Id: uvm_pdaemon.c,v 1.1.2.32 1998/02/06 05:26:30 chs Exp 44 * 45 * 46 * Copyright (c) 1987, 1990 Carnegie-Mellon University. 47 * All rights reserved. 48 * 49 * Permission to use, copy, modify and distribute this software and 50 * its documentation is hereby granted, provided that both the copyright 51 * notice and this permission notice appear in all copies of the 52 * software, derivative works or modified versions, and any portions 53 * thereof, and that both notices appear in supporting documentation. 54 * 55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS" 56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND 57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE. 58 * 59 * Carnegie Mellon requests users of this software to return to 60 * 61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU 62 * School of Computer Science 63 * Carnegie Mellon University 64 * Pittsburgh PA 15213-3890 65 * 66 * any improvements or extensions that they make and grant Carnegie the 67 * rights to redistribute these changes. 68 */ 69 70 /* 71 * uvm_pdaemon.c: the page daemon 72 */ 73 74 #include <sys/param.h> 75 #include <sys/proc.h> 76 #include <sys/systm.h> 77 #include <sys/kernel.h> 78 #include <sys/pool.h> 79 #include <sys/buf.h> 80 #include <sys/vnode.h> 81 #include <sys/mount.h> 82 83 #include <uvm/uvm.h> 84 85 /* 86 * UVMPD_NUMDIRTYREACTS is how many dirty pages the pagedaemon will reactivate 87 * in a pass thru the inactive list when swap is full. the value should be 88 * "small"... if it's too large we'll cycle the active pages thru the inactive 89 * queue too quickly to for them to be referenced and avoid being freed. 90 */ 91 92 #define UVMPD_NUMDIRTYREACTS 16 93 94 95 /* 96 * local prototypes 97 */ 98 99 void uvmpd_scan(void); 100 boolean_t uvmpd_scan_inactive(struct pglist *); 101 void uvmpd_tune(void); 102 103 /* 104 * uvm_wait: wait (sleep) for the page daemon to free some pages 105 * 106 * => should be called with all locks released 107 * => should _not_ be called by the page daemon (to avoid deadlock) 108 */ 109 110 void 111 uvm_wait(const char *wmsg) 112 { 113 int timo = 0; 114 115 /* 116 * check for page daemon going to sleep (waiting for itself) 117 */ 118 119 if (curproc == uvm.pagedaemon_proc) { 120 /* 121 * now we have a problem: the pagedaemon wants to go to 122 * sleep until it frees more memory. but how can it 123 * free more memory if it is asleep? that is a deadlock. 124 * we have two options: 125 * [1] panic now 126 * [2] put a timeout on the sleep, thus causing the 127 * pagedaemon to only pause (rather than sleep forever) 128 * 129 * note that option [2] will only help us if we get lucky 130 * and some other process on the system breaks the deadlock 131 * by exiting or freeing memory (thus allowing the pagedaemon 132 * to continue). for now we panic if DEBUG is defined, 133 * otherwise we hope for the best with option [2] (better 134 * yet, this should never happen in the first place!). 135 */ 136 137 printf("pagedaemon: deadlock detected!\n"); 138 timo = hz >> 3; /* set timeout */ 139 #if defined(DEBUG) 140 /* DEBUG: panic so we can debug it */ 141 panic("pagedaemon deadlock"); 142 #endif 143 } 144 145 uvm_lock_fpageq(); 146 wakeup(&uvm.pagedaemon); /* wake the daemon! */ 147 msleep(&uvmexp.free, &uvm.fpageqlock, PVM | PNORELOCK, wmsg, timo); 148 } 149 150 /* 151 * uvmpd_tune: tune paging parameters 152 * 153 * => called whenever memory is added to (or removed from?) the system 154 * => caller must call with page queues locked 155 */ 156 157 void 158 uvmpd_tune(void) 159 { 160 161 uvmexp.freemin = uvmexp.npages / 30; 162 163 /* between 16k and 512k */ 164 /* XXX: what are these values good for? */ 165 uvmexp.freemin = max(uvmexp.freemin, (16*1024) >> PAGE_SHIFT); 166 #if 0 167 uvmexp.freemin = min(uvmexp.freemin, (512*1024) >> PAGE_SHIFT); 168 #endif 169 170 /* Make sure there's always a user page free. */ 171 if (uvmexp.freemin < uvmexp.reserve_kernel + 1) 172 uvmexp.freemin = uvmexp.reserve_kernel + 1; 173 174 uvmexp.freetarg = (uvmexp.freemin * 4) / 3; 175 if (uvmexp.freetarg <= uvmexp.freemin) 176 uvmexp.freetarg = uvmexp.freemin + 1; 177 178 /* uvmexp.inactarg: computed in main daemon loop */ 179 180 uvmexp.wiredmax = uvmexp.npages / 3; 181 } 182 183 /* 184 * uvm_pageout: the main loop for the pagedaemon 185 */ 186 187 void 188 uvm_pageout(void *arg) 189 { 190 struct uvm_constraint_range constraint; 191 struct uvm_pmalloc *pma; 192 int work_done; 193 int npages = 0; 194 195 /* 196 * ensure correct priority and set paging parameters... 197 */ 198 199 uvm.pagedaemon_proc = curproc; 200 (void) spl0(); 201 uvm_lock_pageq(); 202 npages = uvmexp.npages; 203 uvmpd_tune(); 204 uvm_unlock_pageq(); 205 206 /* 207 * main loop 208 */ 209 210 for (;;) { 211 long size; 212 work_done = 0; /* No work done this iteration. */ 213 214 uvm_lock_fpageq(); 215 216 if (TAILQ_EMPTY(&uvm.pmr_control.allocs)) { 217 msleep(&uvm.pagedaemon, &uvm.fpageqlock, PVM, 218 "pgdaemon", 0); 219 uvmexp.pdwoke++; 220 } 221 222 if ((pma = TAILQ_FIRST(&uvm.pmr_control.allocs)) != NULL) { 223 pma->pm_flags |= UVM_PMA_BUSY; 224 constraint = pma->pm_constraint; 225 } else 226 constraint = no_constraint; 227 228 uvm_unlock_fpageq(); 229 230 /* 231 * now lock page queues and recompute inactive count 232 */ 233 234 uvm_lock_pageq(); 235 if (npages != uvmexp.npages) { /* check for new pages? */ 236 npages = uvmexp.npages; 237 uvmpd_tune(); 238 } 239 240 uvmexp.inactarg = (uvmexp.active + uvmexp.inactive) / 3; 241 if (uvmexp.inactarg <= uvmexp.freetarg) { 242 uvmexp.inactarg = uvmexp.freetarg + 1; 243 } 244 245 /* 246 * Reclaim pages from the buffer cache if possible. 247 */ 248 size = 0; 249 if (pma != NULL) 250 size += pma->pm_size >> PAGE_SHIFT; 251 if (uvmexp.free - BUFPAGES_DEFICIT < uvmexp.freetarg) 252 size += uvmexp.freetarg - uvmexp.free - 253 BUFPAGES_DEFICIT; 254 (void) bufbackoff(&constraint, size * 2); 255 256 /* 257 * Scan if needed to meet our targets. 258 */ 259 if (pma != NULL || 260 ((uvmexp.free - BUFPAGES_DEFICIT) < uvmexp.freetarg) || 261 ((uvmexp.inactive + BUFPAGES_INACT) < uvmexp.inactarg)) { 262 uvmpd_scan(); 263 work_done = 1; /* XXX we hope... */ 264 } 265 266 /* 267 * if there's any free memory to be had, 268 * wake up any waiters. 269 */ 270 uvm_lock_fpageq(); 271 if (uvmexp.free > uvmexp.reserve_kernel || 272 uvmexp.paging == 0) { 273 wakeup(&uvmexp.free); 274 } 275 276 if (pma != NULL) { 277 pma->pm_flags &= ~UVM_PMA_BUSY; 278 if (!work_done) 279 pma->pm_flags |= UVM_PMA_FAIL; 280 if (pma->pm_flags & (UVM_PMA_FAIL | UVM_PMA_FREED)) { 281 pma->pm_flags &= ~UVM_PMA_LINKED; 282 TAILQ_REMOVE(&uvm.pmr_control.allocs, pma, 283 pmq); 284 } 285 wakeup(pma); 286 } 287 uvm_unlock_fpageq(); 288 289 /* 290 * scan done. unlock page queues (the only lock we are holding) 291 */ 292 293 uvm_unlock_pageq(); 294 } 295 /*NOTREACHED*/ 296 } 297 298 299 /* 300 * uvm_aiodone_daemon: main loop for the aiodone daemon. 301 */ 302 303 void 304 uvm_aiodone_daemon(void *arg) 305 { 306 int s, free; 307 struct buf *bp, *nbp; 308 309 uvm.aiodoned_proc = curproc; 310 311 for (;;) { 312 313 /* 314 * Check for done aio structures. If we've got structures to 315 * process, do so. Otherwise sleep while avoiding races. 316 */ 317 mtx_enter(&uvm.aiodoned_lock); 318 while ((bp = TAILQ_FIRST(&uvm.aio_done)) == NULL) 319 msleep(&uvm.aiodoned, &uvm.aiodoned_lock, 320 PVM, "aiodoned", 0); 321 /* Take the list for ourselves. */ 322 TAILQ_INIT(&uvm.aio_done); 323 mtx_leave(&uvm.aiodoned_lock); 324 325 /* 326 * process each i/o that's done. 327 */ 328 329 free = uvmexp.free; 330 while (bp != NULL) { 331 if (bp->b_flags & B_PDAEMON) { 332 uvmexp.paging -= bp->b_bufsize >> PAGE_SHIFT; 333 } 334 nbp = TAILQ_NEXT(bp, b_freelist); 335 s = splbio(); /* b_iodone must by called at splbio */ 336 (*bp->b_iodone)(bp); 337 splx(s); 338 bp = nbp; 339 } 340 uvm_lock_fpageq(); 341 wakeup(free <= uvmexp.reserve_kernel ? &uvm.pagedaemon : 342 &uvmexp.free); 343 uvm_unlock_fpageq(); 344 } 345 } 346 347 348 349 /* 350 * uvmpd_scan_inactive: scan an inactive list for pages to clean or free. 351 * 352 * => called with page queues locked 353 * => we work on meeting our free target by converting inactive pages 354 * into free pages. 355 * => we handle the building of swap-backed clusters 356 * => we return TRUE if we are exiting because we met our target 357 */ 358 359 boolean_t 360 uvmpd_scan_inactive(struct pglist *pglst) 361 { 362 boolean_t retval = FALSE; /* assume we haven't hit target */ 363 int free, result; 364 struct vm_page *p, *nextpg; 365 struct uvm_object *uobj; 366 struct vm_page *pps[MAXBSIZE >> PAGE_SHIFT], **ppsp; 367 int npages; 368 struct vm_page *swpps[MAXBSIZE >> PAGE_SHIFT]; /* XXX: see below */ 369 int swnpages, swcpages; /* XXX: see below */ 370 int swslot; 371 struct vm_anon *anon; 372 boolean_t swap_backed; 373 vaddr_t start; 374 int dirtyreacts; 375 376 /* 377 * note: we currently keep swap-backed pages on a separate inactive 378 * list from object-backed pages. however, merging the two lists 379 * back together again hasn't been ruled out. thus, we keep our 380 * swap cluster in "swpps" rather than in pps (allows us to mix 381 * clustering types in the event of a mixed inactive queue). 382 */ 383 384 /* 385 * swslot is non-zero if we are building a swap cluster. we want 386 * to stay in the loop while we have a page to scan or we have 387 * a swap-cluster to build. 388 */ 389 390 swslot = 0; 391 swnpages = swcpages = 0; 392 free = 0; 393 dirtyreacts = 0; 394 395 for (p = TAILQ_FIRST(pglst); p != NULL || swslot != 0; p = nextpg) { 396 397 /* 398 * note that p can be NULL iff we have traversed the whole 399 * list and need to do one final swap-backed clustered pageout. 400 */ 401 402 uobj = NULL; 403 anon = NULL; 404 405 if (p) { 406 407 /* 408 * update our copy of "free" and see if we've met 409 * our target 410 */ 411 free = uvmexp.free - BUFPAGES_DEFICIT; 412 413 if (free + uvmexp.paging >= uvmexp.freetarg << 2 || 414 dirtyreacts == UVMPD_NUMDIRTYREACTS) { 415 retval = TRUE; 416 417 if (swslot == 0) { 418 /* exit now if no swap-i/o pending */ 419 break; 420 } 421 422 /* set p to null to signal final swap i/o */ 423 p = NULL; 424 } 425 } 426 427 if (p) { /* if (we have a new page to consider) */ 428 429 /* 430 * we are below target and have a new page to consider. 431 */ 432 uvmexp.pdscans++; 433 nextpg = TAILQ_NEXT(p, pageq); 434 435 /* 436 * move referenced pages back to active queue and 437 * skip to next page (unlikely to happen since 438 * inactive pages shouldn't have any valid mappings 439 * and we cleared reference before deactivating). 440 */ 441 442 if (pmap_is_referenced(p)) { 443 uvm_pageactivate(p); 444 uvmexp.pdreact++; 445 continue; 446 } 447 448 /* 449 * the only time we expect to see an ownerless page 450 * (i.e. a page with no uobject and !PQ_ANON) is if an 451 * anon has loaned a page from a uvm_object and the 452 * uvm_object has dropped the ownership. in that 453 * case, the anon can "take over" the loaned page 454 * and make it its own. 455 */ 456 457 /* is page part of an anon or ownerless ? */ 458 if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) { 459 anon = p->uanon; 460 KASSERT(anon != NULL); 461 462 /* 463 * if the page is ownerless, claim it in the 464 * name of "anon"! 465 */ 466 467 if ((p->pg_flags & PQ_ANON) == 0) { 468 KASSERT(p->loan_count > 0); 469 p->loan_count--; 470 atomic_setbits_int(&p->pg_flags, 471 PQ_ANON); 472 /* anon now owns it */ 473 } 474 if (p->pg_flags & PG_BUSY) { 475 uvmexp.pdbusy++; 476 /* someone else owns page, skip it */ 477 continue; 478 } 479 uvmexp.pdanscan++; 480 } else { 481 uobj = p->uobject; 482 KASSERT(uobj != NULL); 483 if (p->pg_flags & PG_BUSY) { 484 uvmexp.pdbusy++; 485 /* someone else owns page, skip it */ 486 continue; 487 } 488 uvmexp.pdobscan++; 489 } 490 491 /* 492 * we now have the page queues locked. 493 * the page is not busy. if the page is clean we 494 * can free it now and continue. 495 */ 496 497 if (p->pg_flags & PG_CLEAN) { 498 if (p->pg_flags & PQ_SWAPBACKED) { 499 /* this page now lives only in swap */ 500 uvmexp.swpgonly++; 501 } 502 503 /* zap all mappings with pmap_page_protect... */ 504 pmap_page_protect(p, VM_PROT_NONE); 505 uvm_pagefree(p); 506 uvmexp.pdfreed++; 507 508 if (anon) { 509 510 /* 511 * an anonymous page can only be clean 512 * if it has backing store assigned. 513 */ 514 515 KASSERT(anon->an_swslot != 0); 516 517 /* remove from object */ 518 anon->an_page = NULL; 519 } 520 continue; 521 } 522 523 /* 524 * this page is dirty, skip it if we'll have met our 525 * free target when all the current pageouts complete. 526 */ 527 528 if (free + uvmexp.paging > uvmexp.freetarg << 2) { 529 continue; 530 } 531 532 /* 533 * this page is dirty, but we can't page it out 534 * since all pages in swap are only in swap. 535 * reactivate it so that we eventually cycle 536 * all pages thru the inactive queue. 537 */ 538 539 KASSERT(uvmexp.swpgonly <= uvmexp.swpages); 540 if ((p->pg_flags & PQ_SWAPBACKED) && 541 uvmexp.swpgonly == uvmexp.swpages) { 542 dirtyreacts++; 543 uvm_pageactivate(p); 544 continue; 545 } 546 547 /* 548 * if the page is swap-backed and dirty and swap space 549 * is full, free any swap allocated to the page 550 * so that other pages can be paged out. 551 */ 552 553 KASSERT(uvmexp.swpginuse <= uvmexp.swpages); 554 if ((p->pg_flags & PQ_SWAPBACKED) && 555 uvmexp.swpginuse == uvmexp.swpages) { 556 557 if ((p->pg_flags & PQ_ANON) && 558 p->uanon->an_swslot) { 559 uvm_swap_free(p->uanon->an_swslot, 1); 560 p->uanon->an_swslot = 0; 561 } 562 if (p->pg_flags & PQ_AOBJ) { 563 uao_dropswap(p->uobject, 564 p->offset >> PAGE_SHIFT); 565 } 566 } 567 568 /* 569 * the page we are looking at is dirty. we must 570 * clean it before it can be freed. to do this we 571 * first mark the page busy so that no one else will 572 * touch the page. we write protect all the mappings 573 * of the page so that no one touches it while it is 574 * in I/O. 575 */ 576 577 swap_backed = ((p->pg_flags & PQ_SWAPBACKED) != 0); 578 atomic_setbits_int(&p->pg_flags, PG_BUSY); 579 UVM_PAGE_OWN(p, "scan_inactive"); 580 pmap_page_protect(p, VM_PROT_READ); 581 uvmexp.pgswapout++; 582 583 /* 584 * for swap-backed pages we need to (re)allocate 585 * swap space. 586 */ 587 588 if (swap_backed) { 589 590 /* 591 * free old swap slot (if any) 592 */ 593 594 if (anon) { 595 if (anon->an_swslot) { 596 uvm_swap_free(anon->an_swslot, 597 1); 598 anon->an_swslot = 0; 599 } 600 } else { 601 uao_dropswap(uobj, 602 p->offset >> PAGE_SHIFT); 603 } 604 605 /* 606 * start new cluster (if necessary) 607 */ 608 609 if (swslot == 0) { 610 swnpages = MAXBSIZE >> PAGE_SHIFT; 611 swslot = uvm_swap_alloc(&swnpages, 612 TRUE); 613 if (swslot == 0) { 614 /* no swap? give up! */ 615 atomic_clearbits_int( 616 &p->pg_flags, 617 PG_BUSY); 618 UVM_PAGE_OWN(p, NULL); 619 continue; 620 } 621 swcpages = 0; /* cluster is empty */ 622 } 623 624 /* 625 * add block to cluster 626 */ 627 628 swpps[swcpages] = p; 629 if (anon) 630 anon->an_swslot = swslot + swcpages; 631 else 632 uao_set_swslot(uobj, 633 p->offset >> PAGE_SHIFT, 634 swslot + swcpages); 635 swcpages++; 636 } 637 } else { 638 639 /* if p == NULL we must be doing a last swap i/o */ 640 swap_backed = TRUE; 641 } 642 643 /* 644 * now consider doing the pageout. 645 * 646 * for swap-backed pages, we do the pageout if we have either 647 * filled the cluster (in which case (swnpages == swcpages) or 648 * run out of pages (p == NULL). 649 * 650 * for object pages, we always do the pageout. 651 */ 652 653 if (swap_backed) { 654 if (p) { /* if we just added a page to cluster */ 655 /* cluster not full yet? */ 656 if (swcpages < swnpages) 657 continue; 658 } 659 660 /* starting I/O now... set up for it */ 661 npages = swcpages; 662 ppsp = swpps; 663 /* for swap-backed pages only */ 664 start = (vaddr_t) swslot; 665 666 /* if this is final pageout we could have a few 667 * extra swap blocks */ 668 if (swcpages < swnpages) { 669 uvm_swap_free(swslot + swcpages, 670 (swnpages - swcpages)); 671 } 672 } else { 673 /* normal object pageout */ 674 ppsp = pps; 675 npages = sizeof(pps) / sizeof(struct vm_page *); 676 /* not looked at because PGO_ALLPAGES is set */ 677 start = 0; 678 } 679 680 /* 681 * now do the pageout. 682 * 683 * for swap_backed pages we have already built the cluster. 684 * for !swap_backed pages, uvm_pager_put will call the object's 685 * "make put cluster" function to build a cluster on our behalf. 686 * 687 * we pass the PGO_PDFREECLUST flag to uvm_pager_put to instruct 688 * it to free the cluster pages for us on a successful I/O (it 689 * always does this for un-successful I/O requests). this 690 * allows us to do clustered pageout without having to deal 691 * with cluster pages at this level. 692 * 693 * note locking semantics of uvm_pager_put with PGO_PDFREECLUST: 694 * IN: locked: page queues 695 * OUT: locked: 696 * !locked: pageqs 697 */ 698 699 uvmexp.pdpageouts++; 700 result = uvm_pager_put(swap_backed ? NULL : uobj, p, 701 &ppsp, &npages, PGO_ALLPAGES|PGO_PDFREECLUST, start, 0); 702 703 /* 704 * if we did i/o to swap, zero swslot to indicate that we are 705 * no longer building a swap-backed cluster. 706 */ 707 708 if (swap_backed) 709 swslot = 0; /* done with this cluster */ 710 711 /* 712 * first, we check for VM_PAGER_PEND which means that the 713 * async I/O is in progress and the async I/O done routine 714 * will clean up after us. in this case we move on to the 715 * next page. 716 * 717 * there is a very remote chance that the pending async i/o can 718 * finish _before_ we get here. if that happens, our page "p" 719 * may no longer be on the inactive queue. so we verify this 720 * when determining the next page (starting over at the head if 721 * we've lost our inactive page). 722 */ 723 724 if (result == VM_PAGER_PEND) { 725 uvmexp.paging += npages; 726 uvm_lock_pageq(); 727 uvmexp.pdpending++; 728 if (p) { 729 if (p->pg_flags & PQ_INACTIVE) 730 nextpg = TAILQ_NEXT(p, pageq); 731 else 732 nextpg = TAILQ_FIRST(pglst); 733 } else { 734 nextpg = NULL; 735 } 736 continue; 737 } 738 739 #ifdef UBC 740 if (result == VM_PAGER_ERROR && 741 curproc == uvm.pagedaemon_proc) { 742 uvm_lock_pageq(); 743 nextpg = TAILQ_NEXT(p, pageq); 744 uvm_pageactivate(p); 745 continue; 746 } 747 #endif 748 749 /* 750 * clean up "p" if we have one 751 */ 752 753 if (p) { 754 /* 755 * the I/O request to "p" is done and uvm_pager_put 756 * has freed any cluster pages it may have allocated 757 * during I/O. all that is left for us to do is 758 * clean up page "p" (which is still PG_BUSY). 759 * 760 * our result could be one of the following: 761 * VM_PAGER_OK: successful pageout 762 * 763 * VM_PAGER_AGAIN: tmp resource shortage, we skip 764 * to next page 765 * VM_PAGER_{FAIL,ERROR,BAD}: an error. we 766 * "reactivate" page to get it out of the way (it 767 * will eventually drift back into the inactive 768 * queue for a retry). 769 * VM_PAGER_UNLOCK: should never see this as it is 770 * only valid for "get" operations 771 */ 772 773 /* relock p's object: page queues not lock yet, so 774 * no need for "try" */ 775 776 #ifdef DIAGNOSTIC 777 if (result == VM_PAGER_UNLOCK) 778 panic("pagedaemon: pageout returned " 779 "invalid 'unlock' code"); 780 #endif 781 782 /* handle PG_WANTED now */ 783 if (p->pg_flags & PG_WANTED) 784 wakeup(p); 785 786 atomic_clearbits_int(&p->pg_flags, PG_BUSY|PG_WANTED); 787 UVM_PAGE_OWN(p, NULL); 788 789 /* released during I/O? Can only happen for anons */ 790 if (p->pg_flags & PG_RELEASED) { 791 KASSERT(anon != NULL); 792 /* 793 * remove page so we can get nextpg, 794 * also zero out anon so we don't use 795 * it after the free. 796 */ 797 anon->an_page = NULL; 798 p->uanon = NULL; 799 800 uvm_anfree(anon); /* kills anon */ 801 pmap_page_protect(p, VM_PROT_NONE); 802 anon = NULL; 803 uvm_lock_pageq(); 804 nextpg = TAILQ_NEXT(p, pageq); 805 /* free released page */ 806 uvm_pagefree(p); 807 } else { /* page was not released during I/O */ 808 uvm_lock_pageq(); 809 nextpg = TAILQ_NEXT(p, pageq); 810 if (result != VM_PAGER_OK) { 811 /* pageout was a failure... */ 812 if (result != VM_PAGER_AGAIN) 813 uvm_pageactivate(p); 814 pmap_clear_reference(p); 815 /* XXXCDC: if (swap_backed) FREE p's 816 * swap block? */ 817 } else { 818 /* pageout was a success... */ 819 pmap_clear_reference(p); 820 pmap_clear_modify(p); 821 atomic_setbits_int(&p->pg_flags, 822 PG_CLEAN); 823 } 824 } 825 826 /* 827 * drop object lock (if there is an object left). do 828 * a safety check of nextpg to make sure it is on the 829 * inactive queue (it should be since PG_BUSY pages on 830 * the inactive queue can't be re-queued [note: not 831 * true for active queue]). 832 */ 833 834 if (nextpg && (nextpg->pg_flags & PQ_INACTIVE) == 0) { 835 nextpg = TAILQ_FIRST(pglst); /* reload! */ 836 } 837 } else { 838 839 /* 840 * if p is null in this loop, make sure it stays null 841 * in the next loop. 842 */ 843 844 nextpg = NULL; 845 846 /* 847 * lock page queues here just so they're always locked 848 * at the end of the loop. 849 */ 850 851 uvm_lock_pageq(); 852 } 853 } 854 return (retval); 855 } 856 857 /* 858 * uvmpd_scan: scan the page queues and attempt to meet our targets. 859 * 860 * => called with pageq's locked 861 */ 862 863 void 864 uvmpd_scan(void) 865 { 866 int free, inactive_shortage, swap_shortage, pages_freed; 867 struct vm_page *p, *nextpg; 868 struct uvm_object *uobj; 869 boolean_t got_it; 870 871 uvmexp.pdrevs++; /* counter */ 872 uobj = NULL; 873 874 /* 875 * get current "free" page count 876 */ 877 free = uvmexp.free - BUFPAGES_DEFICIT; 878 879 #ifndef __SWAP_BROKEN 880 /* 881 * swap out some processes if we are below our free target. 882 * we need to unlock the page queues for this. 883 */ 884 if (free < uvmexp.freetarg) { 885 uvmexp.pdswout++; 886 uvm_unlock_pageq(); 887 uvm_swapout_threads(); 888 uvm_lock_pageq(); 889 } 890 #endif 891 892 /* 893 * now we want to work on meeting our targets. first we work on our 894 * free target by converting inactive pages into free pages. then 895 * we work on meeting our inactive target by converting active pages 896 * to inactive ones. 897 */ 898 899 /* 900 * alternate starting queue between swap and object based on the 901 * low bit of uvmexp.pdrevs (which we bump by one each call). 902 */ 903 904 got_it = FALSE; 905 pages_freed = uvmexp.pdfreed; /* XXX - int */ 906 if ((uvmexp.pdrevs & 1) != 0 && uvmexp.nswapdev != 0) 907 got_it = uvmpd_scan_inactive(&uvm.page_inactive_swp); 908 if (!got_it) 909 got_it = uvmpd_scan_inactive(&uvm.page_inactive_obj); 910 if (!got_it && (uvmexp.pdrevs & 1) == 0 && uvmexp.nswapdev != 0) 911 (void) uvmpd_scan_inactive(&uvm.page_inactive_swp); 912 pages_freed = uvmexp.pdfreed - pages_freed; 913 914 /* 915 * we have done the scan to get free pages. now we work on meeting 916 * our inactive target. 917 */ 918 919 inactive_shortage = uvmexp.inactarg - uvmexp.inactive - BUFPAGES_INACT; 920 921 /* 922 * detect if we're not going to be able to page anything out 923 * until we free some swap resources from active pages. 924 */ 925 926 swap_shortage = 0; 927 if (uvmexp.free < uvmexp.freetarg && 928 uvmexp.swpginuse == uvmexp.swpages && 929 uvmexp.swpgonly < uvmexp.swpages && 930 pages_freed == 0) { 931 swap_shortage = uvmexp.freetarg - uvmexp.free; 932 } 933 934 for (p = TAILQ_FIRST(&uvm.page_active); 935 p != NULL && (inactive_shortage > 0 || swap_shortage > 0); 936 p = nextpg) { 937 nextpg = TAILQ_NEXT(p, pageq); 938 if (p->pg_flags & PG_BUSY) 939 continue; 940 941 /* is page anon owned or ownerless? */ 942 if ((p->pg_flags & PQ_ANON) || p->uobject == NULL) { 943 KASSERT(p->uanon != NULL); 944 945 /* take over the page? */ 946 if ((p->pg_flags & PQ_ANON) == 0) { 947 KASSERT(p->loan_count > 0); 948 p->loan_count--; 949 atomic_setbits_int(&p->pg_flags, PQ_ANON); 950 } 951 } 952 953 /* 954 * skip this page if it's busy. 955 */ 956 957 if ((p->pg_flags & PG_BUSY) != 0) { 958 continue; 959 } 960 961 /* 962 * if there's a shortage of swap, free any swap allocated 963 * to this page so that other pages can be paged out. 964 */ 965 966 if (swap_shortage > 0) { 967 if ((p->pg_flags & PQ_ANON) && p->uanon->an_swslot) { 968 uvm_swap_free(p->uanon->an_swslot, 1); 969 p->uanon->an_swslot = 0; 970 atomic_clearbits_int(&p->pg_flags, PG_CLEAN); 971 swap_shortage--; 972 } 973 if (p->pg_flags & PQ_AOBJ) { 974 int slot = uao_set_swslot(p->uobject, 975 p->offset >> PAGE_SHIFT, 0); 976 if (slot) { 977 uvm_swap_free(slot, 1); 978 atomic_clearbits_int(&p->pg_flags, 979 PG_CLEAN); 980 swap_shortage--; 981 } 982 } 983 } 984 985 /* 986 * deactivate this page if there's a shortage of 987 * inactive pages. 988 */ 989 990 if (inactive_shortage > 0) { 991 pmap_page_protect(p, VM_PROT_NONE); 992 /* no need to check wire_count as pg is "active" */ 993 uvm_pagedeactivate(p); 994 uvmexp.pddeact++; 995 inactive_shortage--; 996 } 997 } 998 } 999